To enhance the automation and spatial accuracy of geological outcrop information extraction, this study proposes a three-dimensional (3D) modeling method oriented toward multi-class semantic boundary representation. High-resolution outcrop images at the centimeter scale were acquired using UAV-based oblique photogrammetry. A modified U-Net model, integrating a self-attention mechanism and a boundary-weighted IoU loss, was developed to achieve fine-grained semantic segmentation of four explicit classes: microbial reef, grain beach dolomite, micritic dolomite, and background. To address the occlusion and label conflict issues inherent in multi-view imaging, a spatially-constrained projection strategy based on Z-buffer depth testing and weighted majority voting was proposed, accurately mapping 2D semantic labels onto 3D photorealistic model surfaces. Following high-precision 3D reconstruction using ContextCapture, a multi-scale semantic visualization platform was ultimately constructed based on Cesium. Experimental results demonstrate that the proposed method increases the boundary-aware semantic segmentation performance (Boundary F1-Score) to 78.64% and limits the mean projection error to within 3.2 cm, significantly outperforming existing mainstream approaches. This research provides an effective pathway for the automated recognition of complex geological structures and the development of digital twin models.
Dong et al. (Thu,) studied this question.